Jukka Zitting created OAK-1032:
----------------------------------
Summary: SegmentMK: One SegmentWriter per SegmentNodeStore
Key: OAK-1032
URL: https://issues.apache.org/jira/browse/OAK-1032
Project: Jackrabbit Oak
Issue Type: Sub-task
Components: core
Reporter: Jukka Zitting
Priority: Minor
Currently each SegmentNodeStoreBranch (and thus Root) instance has it's own
SegmentWriter that it uses to construct new segments. This allows maximum write
concurrency as there's no need for any synchronization at the SegmentWriter
level.
The downside of this solution is that the resulting segments are often very
small (< 100 bytes), which increases write overhead (in TarMK each segment
needs at least 1kB of disk space) and reduces locality of reference (related
content gets scattered across many segments). And concurrent write throughput
is in any case constrained when more than one writer is updating the same
journal.
Thus I think it would be good to instead have just one SegmentWriter per
SegmentNodeStore and have all branches use that shared writer. This requires
synchronization when writing individual records, and can thus lead to extra
contention, but since records are fine-grained and all record writes happen in
memory, the amount of contention should be fairly small for normal workloads
(and for truly heavy concurrent writing, we'd in any case be using multiple
processes across many cluster nodes, in which case this limit wouldn't apply).
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
